X Tutup
.. DO NOT EDIT. .. THIS FILE WAS AUTOMATICALLY GENERATED BY SPHINX-GALLERY. .. TO MAKE CHANGES, EDIT THE SOURCE PYTHON FILE: .. "auto_examples/plot_benchmark_rf.py" .. LINE NUMBERS ARE GIVEN BELOW. .. only:: html .. note:: :class: sphx-glr-download-link-note :ref:`Go to the end ` to download the full example code. .. rst-class:: sphx-glr-example-title .. _sphx_glr_auto_examples_plot_benchmark_rf.py: .. _l-example-benchmark-tree-implementation: Benchmark of TreeEnsemble implementation ======================================== The following example compares the inference time between :epkg:`onnxruntime` and :class:`sklearn.ensemble.RandomForestRegressor`, fow different number of estimators, max depth, and parallelization. It does it for a fixed number of rows and features. import and registration of necessary converters ++++++++++++++++++++++++++++++++++++++++++++++++ .. GENERATED FROM PYTHON SOURCE LINES 15-65 .. code-block:: Python import pickle import os import time from itertools import product import matplotlib.pyplot as plt import numpy import pandas from lightgbm import LGBMRegressor from onnxruntime import InferenceSession, SessionOptions from psutil import cpu_count from sphinx_runpython.runpython import run_cmd from skl2onnx import to_onnx, update_registered_converter from skl2onnx.common.shape_calculator import calculate_linear_regressor_output_shapes from sklearn import set_config from sklearn.ensemble import RandomForestRegressor from tqdm import tqdm from xgboost import XGBRegressor from onnxmltools.convert.xgboost.operator_converters.XGBoost import convert_xgboost def skl2onnx_convert_lightgbm(scope, operator, container): from onnxmltools.convert.lightgbm.operator_converters.LightGbm import ( convert_lightgbm, ) options = scope.get_options(operator.raw_operator) operator.split = options.get("split", None) convert_lightgbm(scope, operator, container) update_registered_converter( LGBMRegressor, "LightGbmLGBMRegressor", calculate_linear_regressor_output_shapes, skl2onnx_convert_lightgbm, options={"split": None}, ) update_registered_converter( XGBRegressor, "XGBoostXGBRegressor", calculate_linear_regressor_output_shapes, convert_xgboost, ) # The following instruction reduces the time spent by scikit-learn # to validate the data. set_config(assume_finite=True) .. GENERATED FROM PYTHON SOURCE LINES 66-68 Machine details +++++++++++++++ .. GENERATED FROM PYTHON SOURCE LINES 68-72 .. code-block:: Python print(f"Number of cores: {cpu_count()}") .. rst-class:: sphx-glr-script-out .. code-block:: none Number of cores: 20 .. GENERATED FROM PYTHON SOURCE LINES 73-75 But this information is not usually enough. Let's extract the cache information. .. GENERATED FROM PYTHON SOURCE LINES 75-82 .. code-block:: Python try: out, err = run_cmd("lscpu") print(out) except Exception as e: print(f"lscpu not available: {e}") .. rst-class:: sphx-glr-script-out .. code-block:: none .. GENERATED FROM PYTHON SOURCE LINES 83-84 Or with the following command. .. GENERATED FROM PYTHON SOURCE LINES 84-87 .. code-block:: Python out, err = run_cmd("cat /proc/cpuinfo") print(out) .. rst-class:: sphx-glr-script-out .. code-block:: none .. GENERATED FROM PYTHON SOURCE LINES 88-90 Fonction to measure inference time ++++++++++++++++++++++++++++++++++ .. GENERATED FROM PYTHON SOURCE LINES 90-121 .. code-block:: Python def measure_inference(fct, X, repeat, max_time=5, quantile=1): """ Run *repeat* times the same function on data *X*. :param fct: fonction to run :param X: data :param repeat: number of times to run :param max_time: maximum time to use to measure the inference :return: number of runs, sum of the time, average, median """ times = [] for _n in range(repeat): perf = time.perf_counter() fct(X) delta = time.perf_counter() - perf times.append(delta) if len(times) < 3: continue if max_time is not None and sum(times) >= max_time: break times.sort() quantile = 0 if (len(times) - quantile * 2) < 3 else quantile if quantile == 0: tt = times else: tt = times[quantile:-quantile] return (len(times), sum(times), sum(tt) / len(tt), times[len(times) // 2]) .. GENERATED FROM PYTHON SOURCE LINES 122-128 Benchmark +++++++++ The following script benchmarks the inference for the same model for a random forest and onnxruntime after it was converted into ONNX and for the following configurations. .. GENERATED FROM PYTHON SOURCE LINES 128-150 .. code-block:: Python small = cpu_count() < 25 if small: N = 1000 n_features = 10 n_jobs = [1, cpu_count() // 2, cpu_count()] n_ests = [10, 20, 30] depth = [4, 6, 8, 10] Regressor = RandomForestRegressor else: N = 100000 n_features = 50 n_jobs = [cpu_count(), cpu_count() // 2, 1] n_ests = [100, 200, 400] depth = [6, 8, 10, 12, 14] Regressor = RandomForestRegressor legend = f"parallel-nf-{n_features}-" # avoid duplicates on machine with 1 or 2 cores. n_jobs = list(sorted(set(n_jobs), reverse=True)) .. GENERATED FROM PYTHON SOURCE LINES 151-152 Benchmark parameters .. GENERATED FROM PYTHON SOURCE LINES 152-156 .. code-block:: Python repeat = 7 # repeat n times the same inference quantile = 1 # exclude extreme times max_time = 5 # maximum number of seconds to spend on one configuration .. GENERATED FROM PYTHON SOURCE LINES 157-158 Data .. GENERATED FROM PYTHON SOURCE LINES 158-251 .. code-block:: Python X = numpy.random.randn(N, n_features).astype(numpy.float32) noise = (numpy.random.randn(X.shape[0]) / (n_features // 5)).astype(numpy.float32) y = X.mean(axis=1) + noise n_train = min(N, N // 3) data = [] couples = list(product(n_jobs, depth, n_ests)) bar = tqdm(couples) cache_dir = "_cache" if not os.path.exists(cache_dir): os.mkdir(cache_dir) for n_j, max_depth, n_estimators in bar: if n_j == 1 and n_estimators > n_ests[0]: # skipping continue # parallelization cache_name = os.path.join( cache_dir, f"nf-{X.shape[1]}-rf-J-{n_j}-E-{n_estimators}-D-{max_depth}.pkl" ) if os.path.exists(cache_name): with open(cache_name, "rb") as f: rf = pickle.load(f) else: bar.set_description(f"J={n_j} E={n_estimators} D={max_depth} train rf") if n_j == 1 and issubclass(Regressor, RandomForestRegressor): rf = Regressor(max_depth=max_depth, n_estimators=n_estimators, n_jobs=-1) rf.fit(X[:n_train], y[:n_train]) rf.n_jobs = 1 else: rf = Regressor(max_depth=max_depth, n_estimators=n_estimators, n_jobs=n_j) rf.fit(X[:n_train], y[:n_train]) with open(cache_name, "wb") as f: pickle.dump(rf, f) bar.set_description(f"J={n_j} E={n_estimators} D={max_depth} ISession") so = SessionOptions() so.intra_op_num_threads = n_j cache_name = os.path.join( cache_dir, f"nf-{X.shape[1]}-rf-J-{n_j}-E-{n_estimators}-D-{max_depth}.onnx" ) if os.path.exists(cache_name): sess = InferenceSession(cache_name, so, providers=["CPUExecutionProvider"]) else: bar.set_description(f"J={n_j} E={n_estimators} D={max_depth} cvt onnx") onx = to_onnx(rf, X[:1]) with open(cache_name, "wb") as f: f.write(onx.SerializeToString()) sess = InferenceSession(cache_name, so, providers=["CPUExecutionProvider"]) onx_size = os.stat(cache_name).st_size # run once to avoid counting the first run bar.set_description(f"J={n_j} E={n_estimators} D={max_depth} predict1") rf.predict(X) sess.run(None, {"X": X}) # fixed data obs = dict( n_jobs=n_j, max_depth=max_depth, n_estimators=n_estimators, repeat=repeat, max_time=max_time, name=rf.__class__.__name__, n_rows=X.shape[0], n_features=X.shape[1], onnx_size=onx_size, ) # baseline bar.set_description(f"J={n_j} E={n_estimators} D={max_depth} predictB") r, t, mean, med = measure_inference(rf.predict, X, repeat=repeat, max_time=max_time) o1 = obs.copy() o1.update(dict(avg=mean, med=med, n_runs=r, ttime=t, name="base")) data.append(o1) # onnxruntime bar.set_description(f"J={n_j} E={n_estimators} D={max_depth} predictO") r, t, mean, med = measure_inference( lambda x, sess=sess: sess.run(None, {"X": x}), X, repeat=repeat, max_time=max_time, ) o2 = obs.copy() o2.update(dict(avg=mean, med=med, n_runs=r, ttime=t, name="ort_")) data.append(o2) .. rst-class:: sphx-glr-script-out .. code-block:: none 0%| | 0/36 [00:00
n_jobs max_depth n_estimators repeat max_time name n_rows n_features onnx_size avg med n_runs ttime
0 20 4 10 7 5 base 1000 10 11460 0.020402 0.019292 7 0.166615
1 20 4 10 7 5 ort_ 1000 10 11460 0.000432 0.000438 7 0.003177
2 20 4 20 7 5 base 1000 10 22145 0.040205 0.042595 7 0.286761
3 20 4 20 7 5 ort_ 1000 10 22145 0.000992 0.000726 7 0.008394
4 20 4 30 7 5 base 1000 10 32536 0.027468 0.028868 7 0.206444
5 20 4 30 7 5 ort_ 1000 10 32536 0.000506 0.000365 7 0.004230
6 20 6 10 7 5 base 1000 10 34530 0.019679 0.017354 7 0.171011
7 20 6 10 7 5 ort_ 1000 10 34530 0.000103 0.000074 7 0.001878
8 20 6 20 7 5 base 1000 10 66529 0.017820 0.017489 7 0.132340
9 20 6 20 7 5 ort_ 1000 10 66529 0.000156 0.000139 7 0.001469
10 20 6 30 7 5 base 1000 10 103420 0.022106 0.021207 7 0.158302
11 20 6 30 7 5 ort_ 1000 10 103420 0.000321 0.000330 7 0.003994
12 20 8 10 7 5 base 1000 10 71350 0.030077 0.026227 7 0.217749
13 20 8 10 7 5 ort_ 1000 10 71350 0.000284 0.000135 7 0.002352
14 20 8 20 7 5 base 1000 10 144167 0.024029 0.021634 7 0.200874
15 20 8 20 7 5 ort_ 1000 10 144167 0.000488 0.000375 7 0.004292
16 20 8 30 7 5 base 1000 10 214110 0.044221 0.039483 7 0.318789
17 20 8 30 7 5 ort_ 1000 10 214110 0.000797 0.000472 7 0.006995
18 20 10 10 7 5 base 1000 10 122048 0.028782 0.031307 7 0.197500
19 20 10 10 7 5 ort_ 1000 10 122048 0.000159 0.000143 7 0.001353
20 20 10 20 7 5 base 1000 10 219442 0.038451 0.034949 7 0.271147
21 20 10 20 7 5 ort_ 1000 10 219442 0.000247 0.000232 7 0.007190
22 20 10 30 7 5 base 1000 10 334072 0.021335 0.019029 7 0.169890
23 20 10 30 7 5 ort_ 1000 10 334072 0.000316 0.000321 7 0.002366
24 10 4 10 7 5 base 1000 10 11679 0.017994 0.017969 7 0.125971
25 10 4 10 7 5 ort_ 1000 10 11679 0.000219 0.000226 7 0.001612
26 10 4 20 7 5 base 1000 10 22656 0.017507 0.017571 7 0.130793
27 10 4 20 7 5 ort_ 1000 10 22656 0.000192 0.000177 7 0.001623
28 10 4 30 7 5 base 1000 10 33412 0.029199 0.029476 7 0.203991
29 10 4 30 7 5 ort_ 1000 10 33412 0.000310 0.000248 7 0.002301
30 10 6 10 7 5 base 1000 10 33727 0.017788 0.017760 7 0.123793
31 10 6 10 7 5 ort_ 1000 10 33727 0.000181 0.000092 7 0.001484
32 10 6 20 7 5 base 1000 10 66894 0.025162 0.029113 7 0.174407
33 10 6 20 7 5 ort_ 1000 10 66894 0.000245 0.000235 7 0.001975
34 10 6 30 7 5 base 1000 10 101960 0.030069 0.029910 7 0.211250
35 10 6 30 7 5 ort_ 1000 10 101960 0.000275 0.000248 7 0.002105
36 10 8 10 7 5 base 1000 10 73532 0.018503 0.018575 7 0.129667
37 10 8 10 7 5 ort_ 1000 10 73532 0.000219 0.000179 7 0.001621
38 10 8 20 7 5 base 1000 10 149551 0.021080 0.019247 7 0.153584
39 10 8 20 7 5 ort_ 1000 10 149551 0.000212 0.000211 7 0.001832
40 10 8 30 7 5 base 1000 10 222210 0.029159 0.029022 7 0.205343
41 10 8 30 7 5 ort_ 1000 10 222210 0.000406 0.000397 7 0.003140
42 10 10 10 7 5 base 1000 10 114799 0.018103 0.017797 7 0.128269
43 10 10 10 7 5 ort_ 1000 10 114799 0.000224 0.000153 7 0.002583
44 10 10 20 7 5 base 1000 10 227708 0.017079 0.017073 7 0.125737
45 10 10 20 7 5 ort_ 1000 10 227708 0.000316 0.000296 7 0.002357
46 10 10 30 7 5 base 1000 10 338925 0.028174 0.027690 7 0.197377
47 10 10 30 7 5 ort_ 1000 10 338925 0.000530 0.000529 7 0.004071
48 1 4 10 7 5 base 1000 10 11168 0.000785 0.000712 7 0.005677
49 1 4 10 7 5 ort_ 1000 10 11168 0.000258 0.000268 7 0.001803
50 1 6 10 7 5 base 1000 10 33216 0.000832 0.000769 7 0.005944
51 1 6 10 7 5 ort_ 1000 10 33216 0.000321 0.000313 7 0.002284
52 1 8 10 7 5 base 1000 10 77275 0.001040 0.001066 7 0.007312
53 1 8 10 7 5 ort_ 1000 10 77275 0.000456 0.000448 7 0.003227
54 1 10 10 7 5 base 1000 10 117218 0.001015 0.001011 7 0.007275
55 1 10 10 7 5 ort_ 1000 10 117218 0.000531 0.000523 7 0.003769


.. GENERATED FROM PYTHON SOURCE LINES 268-270 Plot ++++ .. GENERATED FROM PYTHON SOURCE LINES 270-312 .. code-block:: Python n_rows = len(n_jobs) n_cols = len(n_ests) fig, axes = plt.subplots(n_rows, n_cols, figsize=(4 * n_cols, 4 * n_rows)) fig.suptitle(f"{rf.__class__.__name__}\nX.shape={X.shape}") for n_j, n_estimators in tqdm(product(n_jobs, n_ests)): i = n_jobs.index(n_j) j = n_ests.index(n_estimators) ax = axes[i, j] subdf = df[(df.n_estimators == n_estimators) & (df.n_jobs == n_j)] if subdf.shape[0] == 0: continue piv = subdf.pivot(index="max_depth", columns="name", values=["avg", "med"]) piv.plot(ax=ax, title=f"jobs={n_j}, trees={n_estimators}") ax.set_ylabel(f"n_jobs={n_j}", fontsize="small") ax.set_xlabel("max_depth", fontsize="small") # ratio ax2 = ax.twinx() piv1 = subdf.pivot(index="max_depth", columns="name", values="avg") piv1["speedup"] = piv1.base / piv1.ort_ ax2.plot(piv1.index, piv1.speedup, "b--", label="speedup avg") piv1 = subdf.pivot(index="max_depth", columns="name", values="med") piv1["speedup"] = piv1.base / piv1.ort_ ax2.plot(piv1.index, piv1.speedup, "y--", label="speedup med") ax2.legend(fontsize="x-small") # 1 ax2.plot(piv1.index, [1 for _ in piv1.index], "k--", label="no speedup") for i in range(axes.shape[0]): for j in range(axes.shape[1]): axes[i, j].legend(fontsize="small") fig.tight_layout() fig.savefig(f"{name}-{legend}.png") # plt.show() .. image-sg:: /auto_examples/images/sphx_glr_plot_benchmark_rf_001.png :alt: RandomForestRegressor X.shape=(1000, 10), jobs=20, trees=10, jobs=20, trees=20, jobs=20, trees=30, jobs=10, trees=10, jobs=10, trees=20, jobs=10, trees=30, jobs=1, trees=10 :srcset: /auto_examples/images/sphx_glr_plot_benchmark_rf_001.png :class: sphx-glr-single-img .. rst-class:: sphx-glr-script-out .. code-block:: none 0it [00:00, ?it/s] 4it [00:00, 38.92it/s] 8it [00:00, 32.62it/s] 9it [00:00, 37.54it/s] /home/xadupre/github/onnx-array-api/_doc/examples/plot_benchmark_rf.py:307: UserWarning: No artists with labels found to put in legend. Note that artists whose label start with an underscore are ignored when legend() is called with no argument. axes[i, j].legend(fontsize="small") .. rst-class:: sphx-glr-timing **Total running time of the script:** (0 minutes 13.510 seconds) .. _sphx_glr_download_auto_examples_plot_benchmark_rf.py: .. only:: html .. container:: sphx-glr-footer sphx-glr-footer-example .. container:: sphx-glr-download sphx-glr-download-jupyter :download:`Download Jupyter notebook: plot_benchmark_rf.ipynb ` .. container:: sphx-glr-download sphx-glr-download-python :download:`Download Python source code: plot_benchmark_rf.py ` .. container:: sphx-glr-download sphx-glr-download-zip :download:`Download zipped: plot_benchmark_rf.zip ` .. only:: html .. rst-class:: sphx-glr-signature `Gallery generated by Sphinx-Gallery `_
X Tutup